Change Detection in Multitemporal High Spatial Resolution Remote-Sensing Images Based on Saliency Detection and Spatial Intuitionistic Fuzzy C-Means Clustering

In order to improve the change detection accuracy of multitemporal high spatial resolution remote-sensing (HSRRS) images, a change detection method of multitemporal remote-sensing images based on saliency detection and spatial intuitionistic fuzzy C-means (SIFCM) clustering is proposed. Firstly, the cluster-based saliency cue method is used to obtain the saliency maps of two temporal remote-sensing images; then, the saliency difference is obtained by subtracting the saliency maps of two temporal remote-sensing images; finally, the SIFCM clustering algorithm is used to classify the saliency difference image to obtain the change regions and unchange regions. Two data sets of multitemporal high spatial resolution remote-sensing images are selected as the experimental data. +e detection accuracy of the proposed method is 96.17% and 97.89%. +e results show that the proposed method is a feasible and better performance multitemporal remote-sensing image change detection method.


Introduction
Remote-sensing image change detection is the main mean to identify and monitor the information on land-covered change caused by human activities and natural processes. At present, change detection has been widely used in forest cover monitoring [1,2], land use/land cover monitoring [3,4], disaster assessment [5], urban expansion analysis [6,7], and other fields.
In general, remote-sensing image change detection methods can be divided into the direct comparison method and the postclassification comparison method. e direct comparison method is widely used in practice because of its simplicity and easy implementation [8]. e main steps of the direct comparison method include difference image construction, binary classification, analysis, and evaluation of detection results. At present, algebraic operation and image transformation methods are mainly used to obtain the difference image. e algebraic operation methods include univariate image differencing (UID) [9], vegetation index differencing (VID) [9], image rationing/log rationing (IR/ LR) [9,10], mixed method [11,12], regression [9], and change vector analysis (CVA) [9]. ese methods are widely used in multitemporal remote-sensing image change detection. For example, Huang et al. got the difference image of two temporal remote-sensing images by UID, IR, and mixed method, and then the 2D-OTSU method improved by the firefly algorithm was used to segment the difference image to obtain the change areas [13]; in order to solve the impact of speckle noise on the change detection accuracy of multitemporal SAR images, Gao et al. proposed a modified-logratio (MLR) operator for change detection of targets in forest concealment [14]. Nordberg and Evertson used VID combined with the image regression method to study the impact on decrease of vegetation coverage in the Swedish mountains caused by increased mining [15]. Gong et al. proposed a change detection method for multispectral imagery based on the generative discriminatory classified network (GDCN): the CVA and OTSU methods were first used in the method to obtain the initial change detection result, and then the GDCN was used to obtain the final change detection result [16]. CVA is the most widely used method for the change detection of multitemporal optical remote-sensing images [17,18], and some new methods have evolved into the basis of CVA, such as polar CVA (PCVA) [17], compressed CVA (C 2 VA) [19], robust CVA (RCVA) [20,21], and objectoriented CVA (OCVA) [22]. With the continuous enhancement of commercial HSRRS image acquisition ability, the HSRRS image has gradually become the main data source of change detection. However, with the improvement of spatial resolution, the details of the ground objects become more and more abundant, and the spectral heterogeneity is also increasing. So, the method of constructing the difference image based on pixel-by-pixel comparison faces great challenges [16]. In the method of direct object comparison, it is very difficult to acquire an image object effectively in HSRRS image change detection. e methods of binary classification for different images mainly include clustering [23,24], threshold segmentation [13], Bayesian [25], support vector machine (SVM) [26], and CRF [27]. Among them, clustering is an effective method for unsupervised binary classification of different images. For example, Atasever proposed an unsupervised change detection method based on reconstruction independent component analysis and ABC-K-means clustering: four different functions were first used to achieve difference images, then four different matrices were projected to one feature of reconstruction independent component analysis, and clustering was performed; finally, K-means tuned by artificial bee colony (ABC-K-means) clustering technique has been developed and proposed by following a different strategy in the clustering phase [24]. Ghosh used fuzzy C-means (FCM) and Gustafson Kessel clustering (GKC) to analyze the difference image obtained by CVA and got the change results [28]. Gong et al. obtained the difference image by fusion of the mean ration image and log ration image, and then fuzzy local information C-means (FLICM) was applied to the difference image to obtain the change area [11]. However, there are some problems with the clustering algorithm represented by traditional FCM, such as the objective function is easy to fall into local minimum, the function convergence speed is slow, and it is sensitive to initial value and noise, which affects the accuracy of change detection.
In recent years, with the development of saliency detection technology, it has been widely used in image segmentation, target recognition, image retrieval, and other fields [29]. Now, saliency detection is also used in change detection. For example, aiming at the complex statistical distributions of the change features, Hou et al. proposed the semisupervised building change detection method based on saliency detection combined with the morphological building index (MBI) [30]. Guo and Zhang first got the fused difference image and feature difference image, then cosaliency strategy was used to obtain the saliency maps of two difference images, and FLICM was finally used to cluster and fuse the saliency maps to achieve the final change detection results [31]. e existing change detection methods based on saliency detection mainly rely on the strategy of obtaining the difference image first and then saliency detection. is strategy is based on the difference image to perform saliency detection to obtain the change areas, so the quality of the difference image directly affects the accuracy of subsequent change detection. e high spatial resolution remote-sensing image has the problems of "the same object with different spectra" and "different objects with the same spectrum." Because the spatial features of the image are ignored, the high resolution remote-sensing image obtained by IR, UID, and CVA methods has poor quality of the difference image. In order to solve this problem, this paper combines the advantages of the saliency detection method of image segmentation, proposes a hypothesis: if there are no changes in the two temporal images, the saliency maps should be roughly the same, and if they are different, it should be that the surface has changed and experiments should be designed to verify it. e saliency detection method based on clustering is used in this paper [32]. is method comprehensively considers the spatial cues and contrast cues for the image of the saliency detection and has better performance than the method that only considers the spectral features. So, a novel change detection method of multitemporal high spatial resolution remote-sensing images based on saliency detection and SIFCM clustering is proposed in this paper. We tried a strategy of the first saliency detection and then difference image acquisition in this method; here, the performance of saliency detection is fully utilized to improve the quality of the difference image. In view of the shortcomings of the traditional FCM, after obtaining the difference image, the SIFCM [33] clustering method is introduced to solve the problems of noise sensitivity and lack of spatial information in the difference image analysis. e rest of this paper is organized as follows. e details of the research method are introduced into Section 2, experimental results and analysis are shown in Section 3, and finally, conclusion is given in Section 4.

Research Method
We assume that I 1 and I 2 are remote-sensing images of the same region acquired in T1 and T2 different periods, and I 1 and I 2 are strictly registered. Firstly, the saliency detection method based on clustering is used to detect the saliency of I 1 and I 2 , and the saliency maps S 1 and S 2 of I 1 and I 2 are obtained, respectively; then, the salience difference image (SDI) of I 1 and I 2 is obtained by subtraction of S 1 and S 2 ; next, the SIFCM clustering is performed to obtain change areas and nonchange areas of the SDI by binary classification; finally, a quantitative method is used to evaluate the accuracy of change detection results. e specific flow chart of the proposed method is shown in Figure 1.

Construction of Saliency Difference
Image. If the remotesensing images of the two periods do not change, the two images should have the same and similar saliency map. If there is a change in the remote-sensing images of two periods, the saliency map will also be different. erefore, the saliency difference image can be obtained by the difference operation of the saliency maps of the two images; on this basis, the information of change and unchange in two temporal images is obtained. In order to get the saliency maps of two temporal remote-sensing images, the saliency detection method based on clustering is adopted in this paper [32], and its specific steps are as follows: (1) e remote-sensing image I j is input.
(2) e K-means clustering method is used to divide I j into K-clusters. (1) and (2) are used to calculate the contrast cue (ω c (k)) and spatial cue (ω s (k)) of each cluster, respectively: In formula (1), ‖·‖ 2 is used to calculate the feature space; n i represents the number of pixels in the cluster C i ; N means the number of pixels in the input image; and μ k and μ i are the cluster centers of C k and C i . In formula (2), δ(·) is the Kronecker delta function; o j means the center of the input image I j ; Gaussian kernel Υ(·) is used to calculate the Euclidean distance between the pixel z j i and the image center o j ; variance σ 2 represents the normalized radius of the input image; the normalized coefficients n k means the pixel number of C k ; p j i is the pixel i in the input image I j ; N j represents the j image lattice; M is the number of input images; and b(p j i ) indicates the clustering index.
(4) Contrast cue and spatial cue are fused through the following formula.
In formula (3), ω i (k) represents the saliency clue. (5) e final single image saliency map can be achieved by the following formula: In formula (4), x is the pixel in the input image. e saliency maps (S 1 and S 2 ) of the two temporal remote-sensing images (I 1 and I 2 ) can be obtained through the saliency detection, and then formula (5) is used to gain the SDI by subtraction of the two temporal remote-sensing images of saliency maps:

Spatial Intuitionistic Fuzzy C-Means
Clustering. By using the SIFCM clustering method [33], the obtained SDI is classified into two classes: change class and nonchange class. e specific steps are as follows: (1) v i (i � 1, . . . , c) initial clustering centers are given, we set c � 2 in this paper. (2) Membership (u ij ) is calculated by the following formula: In formula (6), ‖·‖ 2 is used to calculate the Euclidean distance; i � 1, 2; j � 1,. . . ,N; N is the number of pixels in the input image; m is a constant, which is generally taken as 2; and x j means the pixel.  (3) e intuition index is calculated by the following formula: In formula (7), λ is the Lagrange multiplier. (4) e new membership (u ij ′ ) is calculated: (i) In formula (9), NB(x j ) represents the neighborhood pixels of x j , and 5 × 5 equal weight mask is used here. (6) Membership (u ij ″ ) is updated by synthesizing spatial functions: In formula (10), p and q determine the relative weights of initial membership and spatial function, respectively. (7) e new cluster center is calculated according to the following formula: where ε equals to 0.05, the iterative operation is ended; otherwise, step 2 is returned. e maximum number of iterations (max_iter) is set as 50 in this paper.

Experimental Data.
Two groups of multitemporal and high spatial resolution remote-sensing images are selected as the experimental data. Figures 2(a) and 2(b) show the first group of images, and Figures 3(a) and 3(b) show the second group of images. e two sets of data are all from SZTAKI air change benchmark set [34,35], both of which are multitemporal aerial remote-sensing images, and the sizes of images are 952 pixels × 640 pixels. e images have three visible spectral bands: red, green, and blue, and image spatial resolution is 1.5 m. Figure 2(c) is the reference image of the first group, and Figure 3(c) is the reference image of the second group. In Figures 2 and 3, the white area is the actual change area and the black area is the unchanged area. e configuration of the experimental computer is Intel CORE i7-7700HQ, NVIDIA GTX 1060 6G GDDR5, and 8 GB DDR4; the SIFCM is implemented by tensorflow 2.0, and the saliency detection and difference images are implemented by Matlab 2018b.

Comparative Experiment and Result Analysis.
In order to verify the effectiveness and reliability of the proposed method, two groups of comparative experiments were designed. UID, IR, CVA, and SDI proposed in this paper are used to construct difference images in the first group of experiments, and then SIFCM clustering is adopted to analyze the acquired difference images, which are, respectively, recorded as UID-SIFCM, IR-SIFCM, CVA-SIFCM, and SDI-SIFCM (the proposed method). In the second group, PCA-K-means [36], FLICM [37], CWNN [38], and the proposed method are performed for comparative analysis in the second group of experiments. False alarms (FAs), missed alarms (MAs), overall accuracy (OA) [39], Kappa, and F-measure [40] are used as quantitative evaluation indexes to measure the quality of change detection results.
is method involves many parameters, c � 2 and π ij (x), which can be calculated by equation (7). e parameters that may affect the results of the proposed method are mainly ε and max_iter. In order to verify the effect of ε and max_iter on the experimental results, the values of ε will be 0.03, 0.05, and 0.07; max_iter will be 50, 100, and 300, and Figures 2 and 3 will be used as experimental data. e actual number of iterations in Figures 2 and 3 is 18∼19 and 28∼29 times. erefore, max_iter has a value of 50, 100, and 300, and the change detection accuracy is the same. erefore, the number of iterations in this experiment is 50. At the same time, under the premise of determining max_iter � 50, the values of ε are set as 0.03, 0.05, and 0.07. e accuracy does not change from different values, so here we choose ε � 0.05.  Tables 1 and 2, the same conclusion can be drawn.

Comparative Experimental Results
For two sets of images, the proposed method has the highest OA, Kappa, and F-measure values, which are significantly better than the other three methods. e IR-SIFCM is only lower than the proposed method of the OA evaluation index, but there is a serious MA. e MA for the second group of images is as high as 98.34%, and the Kappa and F-measure indicators of the second group of images are lowest; the results of UID-SIFCM and CVA-SIFCM are close to each other in the three evaluation indexes, and the obtained change detection results are also similar; the comprehensive performance is lower than the proposed method. From the visual comparison and quantitative evaluation, we can see that compared with the difference images constructed by the traditional UID, IR and CVA, the difference image constructed by the SDI has the highest quality. e experimental results verify the feasibility and reliability of the proposed method.      group. It can be seen from the visual comparison and quantitative evaluation that the proposed method has some advantages over the traditional method. e experimental results verify the effectiveness and reliability of the proposed method.

Conclusion
Direct comparison is a simple and effective change detection strategy for multitemporal remote-sensing images, but the quality of difference image directly affects the accuracy of subsequent change detection. erefore, a difference image construction method based on saliency difference is proposed in this paper. First, the saliency images of two temporal remote-sensing images are obtained, respectively, and then the SDI is constructed by subtraction of the two saliency maps. In this paper, the SIFCM method is used in the difference image to obtain the change and unchange regions by two kinds of clustering. Two groups of multitemporal aerial remotesensing images are selected as the experimental data, and the overall accuracy is 96.17% and 97.89%, respectively, which proves the validity and reliability of the proposed method. However, the proposed method highly depends on the saliency detection results, and if the detection effect of saliency is not good, it will directly affect the subsequent change detection accuracy. erefore, how to obtain highquality saliency detection results will be the focus of the next research.

Data Availability
e data used to support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest
e authors declare that there are no conflicts of interest regarding the publication of this paper.